Localization ROC Study Research Articles

Detecting lumbar lesions in 99m Tc-MDP SPECT by deep learning: Comparison with physicians.

99m Tc-MDP single-photon emission computed tomography (SPECT) is an established tool for diagnosing lumbar stress, a common cause of low back pain (LBP) in pediatric patients. However, detection of small stress lesions is complicated by the low quality of SPECT, leading to significant interreader variability. The study objectives were to develop an approach based on a deep convolutional neural network (CNN) for detecting lumbar lesions in 99m Tc-MDP scans and to compare its performance to that of physicians in a localization receiver operating characteristic (LROC) study. Sixty-five lesion-absent (LA) 99m Tc-MDP studies performed in pediatric patients for evaluating LBP were retrospectively identified. Projections for an artificial focal lesion were acquired separately by imaging a 99m Tc capillary tube at multiple distances from the collimator. An approach was developed to automatically insert lesions into LA scans to obtain realistic lesion-present (LP) 99m Tc-MDP images while ensuring knowledge of the ground truth. A deep CNN was trained using 2.5D views extracted in LP and LA 99m Tc-MDP image sets. During testing, the CNN was applied in a sliding-window fashion to compute a 3D "heatmap" reporting the probability of a lesion being present at each lumbar location. The algorithm was evaluated using cross-validation on a 99m Tc-MDP test dataset which was also studied by five physicians in a LROC study. LP images in the test set were obtained by incorporating lesions at sites selected by a physician based on clinical likelihood of injury in this population. The deep learning (DL) system slightly outperformed human observers, achieving an area under the LROC curve (AUCLROC ) of 0.830 (95% confidence interval [CI]: [0.758, 0.924]) compared with 0.785 (95% CI: [0.738, 0.830]) for physicians. The AUCLROC for the DL system was higher than that of two readers (difference in AUCLROC [ΔAUCLROC ]=0.049 and 0.053) who participated to the study and slightly lower than that of two other readers (ΔAUCLROC =-0.006 and -0.012). Another reader outperformed DL by a more substantial margin (ΔAUCLROC =-0.053). The DL system provides comparable or superior performance than physicians in localizing small 99m Tc-MDP positive lumbar lesions.

Accounting for anatomical noise in search-capable model observers for planar nuclear imaging.

Model observers intended to predict the diagnostic performance of human observers should account for the effects of both quantum and anatomical noise. We compared the abilities of several visual-search (VS) and scanning Hotelling-type models to account for anatomical noise in a localization receiver operating characteristic (LROC) study involving simulated nuclear medicine images. Our VS observer invoked a two-stage process of search and analysis. The images featured lesions in the prostate and pelvic lymph nodes. Lesion contrast and the geometric resolution and sensitivity of the imaging collimator were the study variables. A set of anthropomorphic mathematical phantoms was imaged with an analytic projector based on eight parallel-hole collimators with different sensitivity and resolution properties. The LROC study was conducted with human observers and the channelized nonprewhitening, channelized Hotelling (CH) and VS model observers. The CH observer was applied in a "background-known-statistically" protocol while the VS observer performed a quasi-background-known-exactly task. Both of these models were applied with and without internal noise in the decision variables. A perceptual search threshold was also tested with the VS observer. The model observers without inefficiencies failed to mimic the average performance trend for the humans. The CH and VS observers with internal noise matched the humans primarily at low collimator sensitivities. With both internal noise and the search threshold, the VS observer attained quantitative agreement with the human observers. Computational efficiency is an important advantage of the VS observer.

An evaluation of iterative reconstruction strategies based on mediastinal lesion detection using hybrid Ga‐67 SPECT images

Using psychophysical studies, the authors have evaluated the effectiveness of various strategies for compensating for physical degradations in SPECT imaging. The particular application was Ga-67-citrate imaging of mediastinal tumors, which was chosen because Ga-67 is a particularly challenging radionuclide for imaging. The test strategies included compensations for nonuniform attenuation, distance-dependent spatial resolution, and scatter applied in various combinations as part of iterative reconstructions with the rescaled block iterative-expectation maximization (RBI-EM) algorithm. The authors also evaluated filtered backprojection reconstructions. Strategies were compared on the basis of human-observer studies of lesion localization and detection accuracy using the localization receiver operating characteristics (LROC) paradigm. These studies involved hybrid images which were obtained by adding the projections of Monte Carlo-simulated lesions to disease-free clinical projection data. The background variability in these images can provide a more realistic assessment of the relative utility of reconstruction strategies than images from anthropomorphic digital phantoms. The clinical datasets were obtained using a GE-VG dual-detector SPECT system with CT-estimated attenuation maps. After determining a target lesion contrast, they conducted pilot LROC studies to obtain a near-optimal set of reconstruction parameters for each strategy, and then conducted the strategy comparison study. The results indicate improved detection accuracy with RBI-EM as more compensations are applied within the reconstruction. The relative rankings of the test strategies agreed in most cases with those of previous studies that employed simulated projections of digital anthropomorphic phantoms, thus confirming the findings of those studies.

Evaluation of Multiclass Model Observers in PET LROC Studies

A localization ROC (LROC) study was conducted to evaluate nonprewhitening matched-filter (NPW) and channelized NPW (CNPW) versions of a multiclass model observer as predictors of human tumor-detection performance with PET images. Target localization is explicitly performed by these model observers. Tumors were placed in the liver, lungs, and background soft tissue of a mathematical phantom, and the data simulation modeled a full-3D acquisition mode. Reconstructions were performed with the FORE+AWOSEM algorithm. The LROC study measured observer performance with 2D images consisting of either coronal, sagittal, or transverse views of the same set of cases. Versions of the CNPW observer based on two previously published difference-of-Gaussian channel models demonstrated good quantitative agreement with human observers. One interpretation of these results treats the CNPW observer as a channelized Hotelling observer with implicit internal noise.

Factors Influencing Lesion Detection in SPECT Lung Images.

An earlier localization ROC (LROC) study that found attenuation correction (AC) degraded the detection of solitary pulmonary nodules (SPN) in hybrid SPECT lung images had several potential shortcomings related to the simulation methods. We sought to address these issues with a revised LROC study. Clinical Tc-99m NeoTect scans acquired with a simultaneous transmission-emission protocol defined the normal cases in a single-slice LROC study. Abnormal cases contained a simulated 1-cm lung lesion. Four rescaled-block-iterative EM (RBI) reconstruction strategies applied: 1) AC, scatter correction (SC), and resolution compensation (RC); 2) AC only; 3) RC only; and 4) no corrections (NC). Images from these strategies underwent 3D Gaussian post-smoothing. Performances were defined by the average area under the LROC curve obtained from three human observers. The strategy ranking in order of decreasing performance was: 1) RBI with RC; 2) RBI with all corrections; 3) RBI with AC; and 4) RBI with no corrections. A multireader-multicase (MRMC) analysis only found significant patient and patient-strategy effects. The conflicting results concerning AC from this study and the previous one may revolve around lesion masking effects, which, by design, were not a factor in the current study.

An optimization of reconstruction parameters and investigation into the impact of photon scatter in /sup 67/Ga SPECT

During /sup 67/Ga citrate SPECT imaging, photon downscatter will occur from higher energy photons into lower energy acquisition windows thus possibly adversely affecting reconstructed image quality. With these additional scattered photons present in projection data, the effect of using more complex reconstruction strategies such as three-dimensional detector response compensation (3-D DRC) and attenuation correction (AC) is unclear. Using a combination of numerical channelized hotelling observers (CHO) and human localization receiver operating characteristics (LROC) studies it has been found that maximum lesion detectability occurs when projection data is reconstructed using two iterations of the rescaled block iterative (RBI) algorithm with both 3-D DRC and patient specific AC, followed by a postreconstruction low-pass 3-D Gaussian filtering with a FWHM of /spl ap/1 cm. These parameters deviate from optimal reconstruction parameters for primary photon-only projection data which finds that maximum lesion detection occurs after 4 RBI iterations with the same 3-D DRC, AC, and postfiltering. As expected, it is also observed that decreased tumor detectability results when scattered photons are present in /sup 67/Ga projection data compared to primary photon only reconstructions (reducing maximum A/sub L/ from 0.79 to 0.58). This decrease has been found to be statistically significant in both human LROC and numerical observer studies, suggesting the need for scatter compensation during /sup 67/Ga citrate imaging.

Effect of block-iterative acceleration on Ga-67 tumor detection in thoracic SPECT

A combination of human localization receiver operating characteristic (LROC) and channelized Hotelling observer (CHO) ROC psychophysical studies were used to investigate how accelerated ordered-subset expectation maximization (OSEM) and rescaled block-iterative (RBI) EM reconstruction affect tumor detection in simulated Ga-67 SPECT images, The tumors were 1-cm-diameter spheres within the chest region of the three-dimensional mathematical cardiac-torso phantom. Previous work with iterative detector resolution compensation showed that eight iterations of the OSEM algorithm with a subset size of eight (16 subsets) offered optimal observer performance. For the LROC study in this paper, the OSEM and RBI algorithms were implemented using subset sizes P and iterations K that satisfied the relation P=K for P=1, 2, 4, and 8. The CHO was applied to reconstruction strategies that deviated from this relation. Results show that using P/spl les/2 penalized observer performance compared to strategies with larger subset sizes. Other researchers have reported on the more stable convergence and noise properties of the RBI algorithm [(Byrne, 1996) and (Lalush and Tsui, 2000)]. In a similar vein, we found that an RBI strategy with a subset size of P produced the same performance as an OSEM strategy with subset size 2P. As neither algorithm displayed a decisive advantage in speed over the other, we conclude that the RBI algorithm is the better choice for accelerating the Ga-67 reconstructions.

LROC analysis of detector-response compensation in SPECT

Localization ROC (LROC) observer studies examined whether detector response compensation (DRC) in ordered-subset, expectation-maximization (OSEM) reconstructions helps in the detection and localization of hot tumors. Simulated gallium (Ga-67) images of the thoracic region were used in the study. The projection data modeled the acquisition of attenuated 93- and 185-keV photons with a medium-energy parallel-hole collimator, but scatter was not modeled. Images were reconstructed with five strategies: 1) OSEM with no DRC; 2) OSEM preceded by restoration filtering; 3) OSEM with iterative DRC; 4) OSEM with an ideal DRC; and 5) filtered backprojection (FBP) with no DRC. All strategies included attenuation correction. There were four LROC studies conducted. In a study using a single tumor activity, the ideal DRC offered the best performance, followed by iterative DRC, restoration filtering, OSEM with no DRC, and FBP. Statistical significance at the 5% level was found between all pairs of strategies except for restoration filtering and OSEM with no DRC. A similar ranking was found for a more realistic study using multiple tumor activities. Additional studies considered the effects of OSEM iteration number and tumor activity on the detection improvement that iterative DRC offered with respect to OSEM with no DRC.

A comparison of human observer LROC and numerical observer ROC for tumor detection in SPECT images

Numerical observers that predict human performance in medical detection tasks can relieve some of the burden of conducting psychophysical studies. Research for this purpose has dealt primarily with receiver operating characteristic (ROC) studies with "signal-known-exactly" (SKE) detection tasks. However, clinical tasks requiring searching for tumors are more closely associated with localization ROC (LROC) studies. The authors have compared performances of humans in a LROC study to performances of a channelized Hotelling observer (CHO) in a SKE ROC study. The task was tumor detection in simulated Ga-67 scans of the chest region. The studies compared different image filters created by varying the dimensionality and cut-off frequency of a 5th-order Butterworth filter. Image reconstruction was by filtered backprojection (FBP) with multiplicative Chang attenuation correction. A total of 35 tumor locations were used. Human LROC results for 4 participants were acquired from a study of 140 images per strategy. The LROC ratings are given as areas under the LROC curve. For the ROC study, 2 constant-Q channel models were used, with parameters determined from a previous comparison of human and CHO performance in a SKE ROC study. The CHO's were applied to 200 noise realizations per location and strategy. The CHO ratings of the filtering strategies are given as areas under the ROC curve averaged over location. Correlation between the human and numerical observers was quantified with Spearman rank correlation tests. Rank correlation coefficients of 0.857 and 0.952 were found. The authors conclude that a ROC study with these constant-Q CHO's may be used to distinguish between considerably superior and inferior strategies, and thus reduce the number of strategies considered by human observers in an LROC study.